I work on the parts of AI systems that fail quietly

I’m a computer science undergraduate at VIT Chennai. Most of what I build sits below the model rather than around it — retrieval, memory, and the control logic that decides what a system does when it isn’t confident.

That focus wasn’t deliberate at first. It came from noticing that the failures which actually hurt are rarely the loud ones. A model that crashes gets fixed. A retriever that returns plausible-but-wrong context, or an agent that loops politely for forty steps without learning anything, keeps running and keeps looking fine.

So the work tends toward systems that know their own limits: retrieval that fuses signals instead of betting on one, memory that preserves history instead of overwriting it, and breakers that trip when progress stops rather than when output stops.

I care about resource constraints for the same reason. A model that assumes infinite memory is a model that hasn’t met a phone.